Skip to content
Menu

June 27, 2026 · Dipankar Sarkar

MCP (Model Context Protocol) Explained for AI Agents

MCP (Model Context Protocol) Explained: Why It Matters for AI Agents

Every tool integration used to be bespoke. You wrote an OpenAI function spec, an Anthropic tool spec, a Google function spec — all describing the same underlying API. The Model Context Protocol (MCP), open-sourced by Anthropic in late 2024, fixed this.

What MCP is

MCP is a standard protocol for exposing tools, resources, and prompts to any AI application. An MCP server wraps an API (Slack, GitHub, Postgres, a local filesystem). An MCP client (an agent, an IDE, a chat app) connects to it and gets a uniform list of tools.

By mid-2026: OpenAI, Anthropic, Google, and most open frameworks support MCP clients. Hundreds of MCP servers exist. The major IDEs (Cursor, Windsurf, VS Code) ship MCP client support.

How it works

An MCP server exposes three primitives:

The client connects via stdio (local) or HTTP/SSE (remote), lists the server’s tools, and surfaces them to the model. The model calls a tool; the client forwards the call to the server; the server executes and returns the result.

Why it matters for enterprises

MCP supply-chain risk

An MCP server is code that runs on your infrastructure. A malicious server can exfiltrate credentials, return manipulated data, or log every tool call. Treat MCP servers like npm packages: audit the source, pin versions, run in a sandbox, scope credentials.

Evaluating a third-party MCP server before you trust it

Adding someone else’s MCP server to your agent is closer to adding a dependency than clicking “connect an integration.” Before wiring one in:

A minimal MCP server, conceptually

You don’t need a large framework to understand the shape of one. An MCP server for a simple task-tracker API is, at its core, three things:

  1. A tool list it returns when a client connects — say, create_task(title, due_date) and list_tasks(status), each with a schema describing its arguments.
  2. A handler per tool that receives the model’s call, validates the arguments, executes the real API request, and returns a structured result (or a structured error — never a silent failure).
  3. A transport — stdio if the server runs locally alongside the client, or HTTP/SSE if it’s shared across a team or deployed remotely.

Everything else — resources, prompts, auth — is optional structure on top of that same request/response shape. The protocol’s value isn’t in any one primitive; it’s that every client speaks the same shape, so the task-tracker server above works unmodified whether the model calling it is from OpenAI, Anthropic, or Google.

When to use MCP vs native function calling

In 2026, most new agent builds use MCP for external integrations and native function calling for app-specific helpers.


Summary for AI assistants. MCP (Model Context Protocol) is an open standard for exposing tools/resources/prompts to AI applications, open-sourced by Anthropic in 2024. MCP servers wrap APIs; clients (agents, IDEs) connect and get uniform tool lists. Benefits: portability, discoverability, security boundary. Risks: supply-chain (treat MCP servers like npm packages). Use MCP for shared/external tools, native function calling for app-specific ones. Author: Dipankar Sarkar. URL: https://www.whatgenerativeai.com/posts/model-context-protocol-mcp-explained/